[X86][SSE41][AVX] Update DPPD/DPPS instruction scheduling on btver2

[avt77]

Summary

This patch includes proper schedule some numbers for SSE 4.1 and AVX instructions on btver2 CPU.

[RKSimon]

I realise this was for SSE41 instructions, but given that its just dot product ops, it might be better to rename it and add the VDPPSY cases as well?

[avt77]

To rename as

[X86][SSE41][AVX] Update instruction scheduling on btver2

Is it OK?

[avt77]

I added VDPPSY instructions scheduling.

[RKSimon]

In D39046#901587, @avt77 wrote:

To rename as

[X86][SSE41][AVX] Update instruction scheduling on btver2

Probably this would be best: "[X86][SSE41][AVX] Update DPPD/DPPS instruction scheduling on btver2"

lib/Target/X86/X86ScheduleBtVer2.td

379

AMD docs says this should be [3,3] - go with that

386

AMD docs says this should be [1,3,3] - go with that

[avt77]

Now all numbers are from AMD docs instead of agner ones.

[RKSimon]

LGTM - Please rebase to fix the SSE41-schedule diffs (test_extractps and test_pextrd) before commit

lib/Target/X86/X86ScheduleBtVer2.td

 
 // FIXME: pipe for system/microcode?
 def : WriteRes<WriteSystem,     [JAny]> { let Latency = 100; }
 def : WriteRes<WriteMicrocoded, [JAny]> { let Latency = 100; }
 def : WriteRes<WriteFence,  [JSAGU]>;
 def : WriteRes<WriteNop, []>;
 
 ////////////////////////////////////////////////////////////////////////////////
+// SSE4.1 instructions.
+////////////////////////////////////////////////////////////////////////////////
+
+def WriteDPPS: SchedWriteRes<[JFPU0, JFPU1]> {
+  let Latency = 11;
+  let ResourceCycles = [3,3];
+  let NumMicroOps = 5;
+}
+def : InstRW<[WriteDPPS], (instregex "(V)?DPPSrri")>;
+
+def WriteDPPSLd: SchedWriteRes<[JLAGU, JFPU0, JFPU1]> {
+  let Latency = 16;
+  let ResourceCycles = [1,3,3];
+  let NumMicroOps = 6;
+}
+def : InstRW<[WriteDPPSLd], (instregex "(V)?DPPSrmi")>;
+
+def WriteDPPD: SchedWriteRes<[JFPU0, JFPU1]> {
+  let Latency = 9;
+  let ResourceCycles = [3,3];
+  let NumMicroOps = 3;
+}
+def : InstRW<[WriteDPPD], (instregex "(V)?DPPDrri")>;
+
+def WriteDPPDLd: SchedWriteRes<[JLAGU, JFPU0, JFPU1]> {
+  let Latency = 14;
+  let ResourceCycles = [1,3,3];
+  let NumMicroOps = 3;
+}
+def : InstRW<[WriteDPPDLd], (instregex "(V)?DPPDrmi")>;
+
+////////////////////////////////////////////////////////////////////////////////
 // SSE4A instructions.
 ////////////////////////////////////////////////////////////////////////////////
 
 def WriteEXTRQ: SchedWriteRes<[JFPU01]> {
   let Latency = 1;
   let ResourceCycles = [1];
 }
 def : InstRW<[WriteEXTRQ], (instregex "EXTRQ")>;
 
 def WriteINSERTQ: SchedWriteRes<[JFPU01]> {
   let Latency = 2;
   let ResourceCycles = [4];
 }
 def : InstRW<[WriteINSERTQ], (instregex "INSERTQ")>;
 
 ////////////////////////////////////////////////////////////////////////////////
 // AVX instructions.
 ////////////////////////////////////////////////////////////////////////////////
 
+def WriteVDPPSY: SchedWriteRes<[JFPU1, JFPU0]> {
+  let Latency = 12;
+  let ResourceCycles = [6, 6];
+  let NumMicroOps = 10;
+}
+def : InstRW<[WriteVDPPSY], (instregex "VDPPSYrr")>;
+
+def WriteVDPPSYLd: SchedWriteRes<[JLAGU, JFPU1, JFPU0]> {
+  let Latency = 17;
+  let ResourceCycles = [1, 6, 6];
+  let NumMicroOps = 11;
+}
+def : InstRW<[WriteVDPPSYLd, ReadAfterLd], (instregex "VDPPSYrm")>;
+
 def WriteFAddY: SchedWriteRes<[JFPU0]> {
   let Latency = 3;
   let ResourceCycles = [2];
 }
 def : InstRW<[WriteFAddY], (instregex "VADD(SUB)?P(S|D)Yrr", "VSUBP(S|D)Yrr")>;
 
 def WriteFAddYLd: SchedWriteRes<[JLAGU, JFPU0]> {
   let Latency = 8;

test/CodeGen/X86/avx-schedule.ll

 ; SKX-LABEL: test_dpps:
 ; SKX:       # BB#0:
 ; SKX-NEXT:    vdpps $7, %ymm1, %ymm0, %ymm0 # sched: [13:1.33]
 ; SKX-NEXT:    vdpps $7, (%rdi), %ymm0, %ymm0 # sched: [20:1.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_dpps:
 ; BTVER2:       # BB#0:
-; BTVER2-NEXT:    vdpps $7, %ymm1, %ymm0, %ymm0 # sched: [3:1.00]
-; BTVER2-NEXT:    vdpps $7, (%rdi), %ymm0, %ymm0 # sched: [8:1.00]
+; BTVER2-NEXT:    vdpps $7, %ymm1, %ymm0, %ymm0 # sched: [12:6.00]
+; BTVER2-NEXT:    vdpps $7, (%rdi), %ymm0, %ymm0 # sched: [17:6.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_dpps:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vdpps $7, %ymm1, %ymm0, %ymm0 # sched: [100:?]
 ; ZNVER1-NEXT:    vdpps $7, (%rdi), %ymm0, %ymm0 # sched: [100:?]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = call <8 x float> @llvm.x86.avx.dp.ps.256(<8 x float> %a0, <8 x float> %a1, i8 7)

test/CodeGen/X86/sse41-schedule.ll

 ; SKX-LABEL: test_dppd:
 ; SKX:       # BB#0:
 ; SKX-NEXT:    vdppd $7, %xmm1, %xmm0, %xmm0 # sched: [9:1.00]
 ; SKX-NEXT:    vdppd $7, (%rdi), %xmm0, %xmm0 # sched: [15:1.00]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_dppd:
 ; BTVER2:       # BB#0:
-; BTVER2-NEXT:    vdppd $7, %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
-; BTVER2-NEXT:    vdppd $7, (%rdi), %xmm0, %xmm0 # sched: [8:1.00]
+; BTVER2-NEXT:    vdppd $7, %xmm1, %xmm0, %xmm0 # sched: [9:3.00]
+; BTVER2-NEXT:    vdppd $7, (%rdi), %xmm0, %xmm0 # sched: [14:3.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_dppd:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vdppd $7, %xmm1, %xmm0, %xmm0 # sched: [100:?]
 ; ZNVER1-NEXT:    vdppd $7, (%rdi), %xmm0, %xmm0 # sched: [100:?]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = call <2 x double> @llvm.x86.sse41.dppd(<2 x double> %a0, <2 x double> %a1, i8 7)

 ; SKX-LABEL: test_dpps:
 ; SKX:       # BB#0:
 ; SKX-NEXT:    vdpps $7, %xmm1, %xmm0, %xmm0 # sched: [13:1.33]
 ; SKX-NEXT:    vdpps $7, (%rdi), %xmm0, %xmm0 # sched: [19:1.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_dpps:
 ; BTVER2:       # BB#0:
-; BTVER2-NEXT:    vdpps $7, %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
-; BTVER2-NEXT:    vdpps $7, (%rdi), %xmm0, %xmm0 # sched: [8:1.00]
+; BTVER2-NEXT:    vdpps $7, %xmm1, %xmm0, %xmm0 # sched: [11:3.00]
+; BTVER2-NEXT:    vdpps $7, (%rdi), %xmm0, %xmm0 # sched: [16:3.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_dpps:
 ; ZNVER1:       # BB#0:
 ; ZNVER1-NEXT:    vdpps $7, %xmm1, %xmm0, %xmm0 # sched: [100:?]
 ; ZNVER1-NEXT:    vdpps $7, (%rdi), %xmm0, %xmm0 # sched: [100:?]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = call <4 x float> @llvm.x86.sse41.dpps(<4 x float> %a0, <4 x float> %a1, i8 7)
   %2 = load <4 x float>, <4 x float> *%a2, align 16
   %3 = call <4 x float> @llvm.x86.sse41.dpps(<4 x float> %1, <4 x float> %2, i8 7)
   ret <4 x float> %3
 }
 declare <4 x float> @llvm.x86.sse41.dpps(<4 x float>, <4 x float>, i8) nounwind readnone
 
-define i32 @test_extractps(<4 x float> %a0, i32 *%a1) {
-; GENERIC-LABEL: test_extractps:
-; GENERIC:       # BB#0:
-; GENERIC-NEXT:    extractps $3, %xmm0, %eax # sched: [3:1.00]
-; GENERIC-NEXT:    extractps $1, %xmm0, (%rdi) # sched: [5:1.00]
-; GENERIC-NEXT:    retq # sched: [1:1.00]
-;
-; SLM-LABEL: test_extractps:
-; SLM:       # BB#0:
-; SLM-NEXT:    extractps $3, %xmm0, %eax # sched: [1:1.00]
-; SLM-NEXT:    extractps $1, %xmm0, (%rdi) # sched: [4:2.00]
-; SLM-NEXT:    retq # sched: [4:1.00]
-;
-; SANDY-LABEL: test_extractps:
-; SANDY:       # BB#0:
-; SANDY-NEXT:    vextractps $3, %xmm0, %eax # sched: [3:1.00]
-; SANDY-NEXT:    vextractps $1, %xmm0, (%rdi) # sched: [5:1.00]
-; SANDY-NEXT:    retq # sched: [1:1.00]
-;
-; HASWELL-LABEL: test_extractps:
-; HASWELL:       # BB#0:
-; HASWELL-NEXT:    vextractps $3, %xmm0, %eax # sched: [2:1.00]
-; HASWELL-NEXT:    vextractps $1, %xmm0, (%rdi) # sched: [1:1.00]
-; HASWELL-NEXT:    retq # sched: [2:1.00]
-;
-; BROADWELL-LABEL: test_extractps:
-; BROADWELL:       # BB#0:
-; BROADWELL-NEXT:    vextractps $3, %xmm0, %eax # sched: [2:1.00]
-; BROADWELL-NEXT:    vextractps $1, %xmm0, (%rdi) # sched: [1:1.00]
-; BROADWELL-NEXT:    retq # sched: [2:1.00]
-;
-; SKYLAKE-LABEL: test_extractps:
-; SKYLAKE:       # BB#0:
-; SKYLAKE-NEXT:    vextractps $3, %xmm0, %eax # sched: [3:1.00]
-; SKYLAKE-NEXT:    vextractps $1, %xmm0, (%rdi) # sched: [2:1.00]
-; SKYLAKE-NEXT:    retq # sched: [7:1.00]
-;
-; SKX-LABEL: test_extractps:
-; SKX:       # BB#0:
-; SKX-NEXT:    vextractps $3, %xmm0, %eax # sched: [3:1.00]
-; SKX-NEXT:    vextractps $1, %xmm0, (%rdi) # sched: [2:1.00]
-; SKX-NEXT:    retq # sched: [7:1.00]
-;
-; BTVER2-LABEL: test_extractps:
-; BTVER2:       # BB#0:
-; BTVER2-NEXT:    vextractps $3, %xmm0, %eax # sched: [1:0.50]
-; BTVER2-NEXT:    vextractps $1, %xmm0, (%rdi) # sched: [6:1.00]
-; BTVER2-NEXT:    retq # sched: [4:1.00]
-;
-; ZNVER1-LABEL: test_extractps:
-; ZNVER1:       # BB#0:
-; ZNVER1-NEXT:    vextractps $3, %xmm0, %eax # sched: [2:2.00]
-; ZNVER1-NEXT:    vextractps $1, %xmm0, (%rdi) # sched: [5:2.50]
-; ZNVER1-NEXT:    retq # sched: [1:0.50]
-  %1 = extractelement <4 x float> %a0, i32 3
-  %2 = extractelement <4 x float> %a0, i32 1
-  %3 = bitcast float %1 to i32
-  %4 = bitcast float %2 to i32
-  store i32 %4, i32 *%a1
-  ret i32 %3
-}
-
 define <4 x float> @test_insertps(<4 x float> %a0, <4 x float> %a1, float *%a2) {
 ; GENERIC-LABEL: test_insertps:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    insertps {{.*#+}} xmm0 = zero,xmm1[0],xmm0[2,3] sched: [1:1.00]
 ; GENERIC-NEXT:    insertps {{.*#+}} xmm0 = xmm0[0,1,2],mem[0] sched: [7:1.00]
 ; GENERIC-NEXT:    retq # sched: [1:1.00]
 ;
 ; SLM-LABEL: test_insertps:
; ZNVER1-NEXT:    retq # sched: [1:0.50]
   store i8 %2, i8 *%a1
   %3 = zext i8 %1 to i32
   ret i32 %3
 }
 
 define i32 @test_pextrd(<4 x i32> %a0, i32 *%a1) {
 ; GENERIC-LABEL: test_pextrd:
 ; GENERIC:       # BB#0:
-; GENERIC-NEXT:    paddd %xmm0, %xmm0 # sched: [1:0.50]
-; GENERIC-NEXT:    pextrd $3, %xmm0, %eax # sched: [3:1.00]
-; GENERIC-NEXT:    pextrd $1, %xmm0, (%rdi) # sched: [5:1.00]
+; GENERIC-NEXT:    extractps $3, %xmm0, %eax # sched: [3:1.00]
+; GENERIC-NEXT:    extractps $1, %xmm0, (%rdi) # sched: [5:1.00]
 ; GENERIC-NEXT:    retq # sched: [1:1.00]
 ;
 ; SLM-LABEL: test_pextrd:
 ; SLM:       # BB#0:
-; SLM-NEXT:    paddd %xmm0, %xmm0 # sched: [1:0.50]
-; SLM-NEXT:    pextrd $3, %xmm0, %eax # sched: [1:1.00]
-; SLM-NEXT:    pextrd $1, %xmm0, (%rdi) # sched: [4:2.00]
+; SLM-NEXT:    extractps $3, %xmm0, %eax # sched: [1:1.00]
+; SLM-NEXT:    extractps $1, %xmm0, (%rdi) # sched: [4:2.00]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-LABEL: test_pextrd:
 ; SANDY:       # BB#0:
-; SANDY-NEXT:    vpaddd %xmm0, %xmm0, %xmm0 # sched: [1:0.50]
-; SANDY-NEXT:    vpextrd $3, %xmm0, %eax # sched: [3:1.00]
-; SANDY-NEXT:    vpextrd $1, %xmm0, (%rdi) # sched: [5:1.00]
+; SANDY-NEXT:    vextractps $3, %xmm0, %eax # sched: [3:1.00]
+; SANDY-NEXT:    vextractps $1, %xmm0, (%rdi) # sched: [5:1.00]
 ; SANDY-NEXT:    retq # sched: [1:1.00]
 ;
 ; HASWELL-LABEL: test_pextrd:
 ; HASWELL:       # BB#0:
-; HASWELL-NEXT:    vpaddd %xmm0, %xmm0, %xmm0 # sched: [1:0.50]
-; HASWELL-NEXT:    vpextrd $3, %xmm0, %eax # sched: [2:1.00]
-; HASWELL-NEXT:    vpextrd $1, %xmm0, (%rdi) # sched: [1:1.00]
+; HASWELL-NEXT:    vextractps $3, %xmm0, %eax # sched: [2:1.00]
+; HASWELL-NEXT:    vextractps $1, %xmm0, (%rdi) # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [2:1.00]
 ;
 ; BROADWELL-LABEL: test_pextrd:
 ; BROADWELL:       # BB#0:
-; BROADWELL-NEXT:    vpaddd %xmm0, %xmm0, %xmm0 # sched: [1:0.50]
-; BROADWELL-NEXT:    vpextrd $3, %xmm0, %eax # sched: [2:1.00]
-; BROADWELL-NEXT:    vpextrd $1, %xmm0, (%rdi) # sched: [1:1.00]
+; BROADWELL-NEXT:    vextractps $3, %xmm0, %eax # sched: [2:1.00]
+; BROADWELL-NEXT:    vextractps $1, %xmm0, (%rdi) # sched: [1:1.00]
 ; BROADWELL-NEXT:    retq # sched: [2:1.00]
 ;
 ; SKYLAKE-LABEL: test_pextrd:
 ; SKYLAKE:       # BB#0:
-; SKYLAKE-NEXT:    vpaddd %xmm0, %xmm0, %xmm0 # sched: [1:0.33]
-; SKYLAKE-NEXT:    vpextrd $3, %xmm0, %eax # sched: [3:1.00]
-; SKYLAKE-NEXT:    vpextrd $1, %xmm0, (%rdi) # sched: [2:1.00]
+; SKYLAKE-NEXT:    vextractps $3, %xmm0, %eax # sched: [3:1.00]
+; SKYLAKE-NEXT:    vextractps $1, %xmm0, (%rdi) # sched: [2:1.00]
 ; SKYLAKE-NEXT:    retq # sched: [7:1.00]
 ;
 ; SKX-LABEL: test_pextrd:
 ; SKX:       # BB#0:
-; SKX-NEXT:    vpaddd %xmm0, %xmm0, %xmm0 # sched: [1:0.33]
-; SKX-NEXT:    vpextrd $3, %xmm0, %eax # sched: [3:1.00]
-; SKX-NEXT:    vpextrd $1, %xmm0, (%rdi) # sched: [2:1.00]
+; SKX-NEXT:    vextractps $3, %xmm0, %eax # sched: [3:1.00]
+; SKX-NEXT:    vextractps $1, %xmm0, (%rdi) # sched: [2:1.00]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-LABEL: test_pextrd:
 ; BTVER2:       # BB#0:
-; BTVER2-NEXT:    vpaddd %xmm0, %xmm0, %xmm0 # sched: [1:0.50]
-; BTVER2-NEXT:    vpextrd $3, %xmm0, %eax # sched: [1:0.50]
-; BTVER2-NEXT:    vpextrd $1, %xmm0, (%rdi) # sched: [6:1.00]
+; BTVER2-NEXT:    vextractps $3, %xmm0, %eax # sched: [1:0.50]
+; BTVER2-NEXT:    vextractps $1, %xmm0, (%rdi) # sched: [6:1.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-LABEL: test_pextrd:
 ; ZNVER1:       # BB#0:
-; ZNVER1-NEXT:    vpaddd %xmm0, %xmm0, %xmm0 # sched: [1:0.25]
-; ZNVER1-NEXT:    vpextrd $3, %xmm0, %eax # sched: [1:0.25]
-; ZNVER1-NEXT:    vpextrd $1, %xmm0, (%rdi) # sched: [8:1.00]
-; ZNVER1-NEXT:    retq # sched: [1:0.50]
-  %1 = add <4 x i32> %a0, %a0
-  %2 = extractelement <4 x i32> %1, i32 3
-  %3 = extractelement <4 x i32> %1, i32 1
-  store i32 %3, i32 *%a1
-  ret i32 %2
+; ZNVER1-NEXT:    vextractps $3, %xmm0, %eax # sched: [2:2.00]
+; ZNVER1-NEXT:    vextractps $1, %xmm0, (%rdi) # sched: [5:2.50]
+; ZNVER1-NEXT:    retq # sched: [1:0.50]
+  %1 = extractelement <4 x i32> %a0, i32 3
+  %2 = extractelement <4 x i32> %a0, i32 1
+  store i32 %2, i32 *%a1
+  ret i32 %1
 }
 
 define i64 @test_pextrq(<2 x i64> %a0, <2 x i64> %a1, i64 *%a2) {
 ; GENERIC-LABEL: test_pextrq:
 ; GENERIC:       # BB#0:
 ; GENERIC-NEXT:    pextrq $1, %xmm0, %rax # sched: [3:1.00]
 ; GENERIC-NEXT:    pextrq $1, %xmm0, (%rdi) # sched: [5:1.00]
 ; GENERIC-NEXT:    retq # sched: [1:1.00]